ELMo-CNN-BiGRU双通道文本情感分类模型

文本情感分类通过对带有情感色彩的主观性文本进行分析和推理，帮助用户更好地做出判断与决策。针对传统情感分类模型难以根据上下文信息调整词向量的问题，提出一种双通道文本情感分类模型。利用ELMo和Glove预训练模型分别生成动态和静态词向量，通过堆叠嵌入2种词向量生成输入向量。采用自注意力机制处理输入向量，计算内部的词依赖关系。构建融合卷积神经网络（CNN）和双向门控递归单元（BiGRU）的双通道神经网络结构，同时获取文本局部特征和全局特征。最终将双通道处理结果进行拼接，经过全连接层处理后输入分类器获得文本情感分类结果。实验结果表明，与同类情感分类模型中性能较优的H-BiGRU模型相比，ELMo-CNN-BiGRU模型在IMDB、yelp和sentiment140数据集上的准确率和F1值分别提升了2.42、1.98、2.52和2.40、1.94、2.43个百分点，具有更好的短文本情感分类效果和稳定性。

ELMo-CNN-BiGRU Dual-Channel Text Sentiment Classification Model

Text sentiment classification helps users make better decisions by analyzing and reasoning subjective texts with emotional colors.Addressing the difficulty in adjusting the word vector according to the context information in traditional sentiment classification models, a dual-channel text sentiment classification method is proposed.To begin, pretrained ELMo and Glove models are used to generate dynamic and static word vectors, respectively.The input vector is generated by stacking and embedding two-word vectors.Second, the self-attention mechanism is used to process the input vector and calculate the internal word dependencies.The dual-channel neural network structure is constructed by a Convolutional Neural Network (CNN) and Bi-directional Gated Recurrent Unit(BiGRU).The local and global features of the text can be obtained simultaneously.Finally, the dual-channel processing results are spliced, processed by the fully connected layer, and sent to the classifier.The classification results can be obtained.The results show that, compared with the H-BiGRU model with the best performance among contrastive sentiment classification models, the accuracy of the proposed ELMo-CNN-BiGRU model on the IMDB, yelp, and sentiment140 datasets improved by 2.42, 1.98, 2.52, respectively, and the F1 value improved by 2.40, 1.94, 2.43 percentage points, respectively.It achieves a better sentiment classification effect and stability for short texts.